[Paper Review] ML for Flood Forecasting at Scale

Sella Nevo, Anisimov, Vova+12 more|arXiv (Cornell University)|Jan 28, 2019

Flood Risk Assessment and ManagementEnvironmental Science17 references19 citations

TL;DR

This paper proposes using machine learning (ML) to scale accurate, real-time riverine flood forecasting by overcoming limitations in manual calibration, data scarcity, and computational costs. It leverages transfer and multitask learning across global basins to improve model generalization and performance, demonstrating high accuracy (90% recall, 75% precision) in a pilot hydro-dynamic model in Patna, India.

ABSTRACT

Effective riverine flood forecasting at scale is hindered by a multitude of factors, most notably the need to rely on human calibration in current methodology, the limited amount of data for a specific location, and the computational difficulty of building continent/global level models that are sufficiently accurate. Machine learning (ML) is primed to be useful in this scenario: learned models often surpass human experts in complex high-dimensional scenarios, and the framework of transfer or multitask learning is an appealing solution for leveraging local signals to achieve improved global performance. We propose to build on these strengths and develop ML systems for timely and accurate riverine flood prediction.

Motivation & Objective

• Address the critical gap in scalable, accurate riverine flood forecasting, especially in data-scarce developing regions.
• Overcome the reliance on manual, time-intensive calibration in hydrologic models that limits scalability and generalization.
• Leverage large-scale, multi-source data (satellite, in-situ, hydrodynamic simulations) to train robust ML models for global flood prediction.
• Develop ML-based solutions for key components of flood forecasting, including discharge estimation, hydrodynamic modeling, and hybrid physics-ML systems.
• Catalyze broader academic and operational collaboration by releasing curated, open data and operational models for the ML community.

Proposed method

• Apply transfer and multitask learning to share knowledge across multiple river basins, improving model performance with limited local data.
• Integrate ML with physics-based hydrodynamic models to reduce computational costs while maintaining accuracy.
• Use remote sensing data (e.g., satellite imagery) to estimate river discharge where in-situ gauges are sparse or absent.
• Train ML models on historical flood events and real-time measurements to predict flood extent and water levels at 300m spatial resolution.
• Deploy a pilot operational system in Patna, India, using real-time data and ML-enhanced simulations to generate inundation maps.
• Combine open data from diverse sources (satellite, gauges, hydrodynamic models) into a unified, scalable framework for model training and evaluation.

Experimental results

Research questions

• RQ1 Can transfer and multitask learning significantly improve flood prediction accuracy in data-scarce basins?
• RQ2 To what extent can ML reduce computational costs in physics-based hydrodynamic modeling without sacrificing accuracy?
• RQ3 How effective is ML in estimating river discharge from satellite data when in-situ measurements are limited?
• RQ4 Can hybrid physics-ML models outperform purely physics-based or purely data-driven models in flood forecasting?
• RQ5 What is the impact of large-scale, multi-basin training on model generalization and operational reliability?

Key findings

• The pilot ML-enhanced hydrodynamic model in Patna, India, achieved over 90% recall and over 75% precision in flood extent prediction at 300m spatial resolution.
• ML-based models demonstrated the ability to generalize across basins by leveraging shared hydrological principles through transfer learning.
• Remote discharge estimation using satellite data showed promise as a viable alternative where in-situ gauges are unavailable.
• The integration of ML with physics-based models reduced computational costs and improved robustness to input errors.
• The system successfully produced real-time flood alerts during the 2018 monsoon season, validating operational feasibility.
• The approach enables scalable, continent-level flood forecasting by overcoming limitations of manual calibration and data scarcity.